LLM Models

A more affordable April 2026 model for coding and professional work, combining advances in reasoning, coding, and agentic workflows. Half the per-token price of GPT-5.5 with a 1M-token context window and multimodal (text + vision) input.

The smaller of OpenAI's two open-weight models (August 2025, Apache 2.0), a 21B-parameter mixture-of-experts model with 3.6B active parameters that delivers results similar to o3-mini and runs on edge devices with just 16GB of memory.

OpenAI's first open-weight language model since GPT-2 (August 2025), released under Apache 2.0. A 117B-parameter mixture-of-experts model with 5.1B active parameters per token, it reaches near-parity with o4-mini on core reasoning benchmarks while running on a single 80GB GPU thanks to MXFP4 quantization.

OpenAI's small reasoning model (January 2025) optimized for STEM - science, math, and coding - at low cost and latency. o3-mini supports tool use and adjustable reasoning effort (low/medium/high) with a 200K context window, and replaced o1-mini as the default small reasoner.

OpenAI's o4-mini (April 2025), a fast, cost-efficient reasoning model that halves the price of o3-mini while matching it on most reasoning benchmarks. It supports much higher usage limits than o3 - a strong high-throughput reasoning option - and can reason with images.

OpenAI's o3 reasoning model (April 2025), which replaced o1 as the primary reasoning model at an ~87% lower price. o3 set state-of-the-art results on Codeforces, SWE-bench, and MMMU, reasons with images, and makes ~20% fewer major errors than o1 on hard real-world tasks.

The fast, affordable model of OpenAI's GPT-5.6 generation (released July 2026), for lower-cost everyday work like summarization, drafting, and routine automation. Luna offers strong Terminal-Bench 2.1 performance at the family's lowest per-token price.

The balanced model of OpenAI's GPT-5.6 generation (released July 2026), tuned for high-volume business work like customer support, internal tools, and document analysis. Terra offers competitive Terminal-Bench 2.1 performance at roughly half GPT-5.5's price.

OpenAI's flagship in the GPT-5.6 generation (released July 2026), for the hardest problems including complex coding and security research. Sol set a new state of the art on Terminal-Bench 2.1, with an 'ultra' subagent mode for the hardest command-line tasks. GPT-5.6 fixes three tiers - Sol, Terra, Luna - and floats the generation number, signaling more frequent capability upgrades.

The premium reasoning tier of GPT-5.5 (April 2026), for the hardest agentic, coding, and knowledge-work tasks. Shares GPT-5.5's 1M-token context window and multimodal (text + vision) input, with extended reasoning for maximum accuracy.

OpenAI's flagship as of April 2026 and its first fully retrained base model since GPT-4.5, built agentic-first. GPT-5.5 is OpenAI's strongest agentic coding model and the engine behind Codex, excelling at coding, computer use, knowledge work, and multi-step task completion. Features a 1M-token context window and multimodal (text + vision) input.